Thanks for responding.
I managed to resolve the problem last Friday; I had a single datasource for each file, instead of one big datasource for all the files. The reading of the one or two HDFS blocks within each datasource was then distributed to a small percentage of slots (let's
say ~10%). Some Beam runner-specific knowledge for Flink I did not yet have.
Using a single TextIO for all the files allowed Flink to use all available parallelism. It did require some hoops to jump through, as there was metadata associated with each file, each PCollection, which was harder to match to each file in one big PCollection.
From: Aljoscha Krettek <aljoscha@xxxxxxxxxx>
Sent: 13 March 2018 18:29:52
Subject: Re: HDFS data locality and distribution, Flink
There should be no data-locality awareness with Beam on Flink because there are no APIs in Beam that Flink could use to schedule tasks with awareness. It seems it just happens that the readers are distributed as they are.
Are the files roughly of equal size?
On 12. Mar 2018, at 05:50, Reinier Kip <rkip@xxxxxxx
Beam 2.1, Flink 1.3.
I'm trying to batch-process 30-ish files from HDFS, but I see that data is distributed very badly across slots. 4 out of 32 slots get 4/5ths of the data, another 3 slots get about 1/5th and a last slot just a few records. This probably triggers disk spillover
on these slots and slows down the job immensely. The data has many many unique keys and processing could be done in a highly parallel manner. From what I understand, HDFS data locality governs which splits are assigned to which subtask.
- I'm running a Beam on Flink on YARN pipeline.
- I'm reading 30-ish files, whose records are later grouped by their millions of unique keys.
- For now, I have 8 task managers by 4 slots. Beam sets all subtasks to have 32 parallelism.
- Data seems to be localised to 9 out of the 32 slots, 3 out of the 8 task managers.
Does the statement of input split assignment ring true? Is the fact that data isn't redistributed an effort from Flink to have high data locality, even if this means disk spillover for a few slots/tms and idleness for others? Is there any use for parallelism
if work isn't distributed anyway?
Thanks for your time, Reinier